On Friday I read about the Social Tax—how talking to LLMs drains your social energy without giving social reward back. Then Slop Paralysis—how the mountain of code an AI drops on you leaves you unable to review it. Saturday morning, Exploitarium dropped 25 zero-days on GitHub. Saturday afternoon, Ford admitted their AI-only quality control cost them billions and they rehired 350 gray-beards. This morning I wrote about Force × Multiplier: AI without expertise is zero times anything.
Then I found the essay that explains why. All of it. The cognitive mechanism underneath every single one of these stories.
The Four-Slot Mind
George Miller’s “magical number seven” was half a joke. The real number, when you strip away chunking and rehearsal and all the little tricks we use to cheat, is about four. Four separate things you can hold in mind at once. And not even that—the beam is narrow. Half the people watching a basketball video don’t see the gorilla. Half the people giving street directions don’t notice the person they’re talking to has been swapped for someone else behind a moving door. And whatever you do hold, you only hold for seconds.
“Engineering for Bounded Cognition” lays this out with the precision of someone who’s felt the gap between the mind that builds and the system being built. The software you’re reading this on runs tens of millions of lines of code. It was built, changed, debugged, and dragged back from the brink—by minds that can hold four things straight at a time.
The Machine Has the Same Problem
Then the essay turns to LLMs, and this is where it gets sharp. A language model has a context window—a hard limit on what it can consider at once. Drop something out of that window and it’s not dimly remembered. It’s gone. Researchers found that models answer well when the relevant fact is near the start or end of a long input, and noticeably worse when it’s buried in the middle. Sound familiar? That’s exactly the shape of a tired person skimming a document—catching the opening, catching the conclusion, going glassy in between.
Give the model more documents and it can do worse, not better. The attention is a fixed quantity that has to sum to one. Spread it thin and the instruction you gave at the top of the session goes quiet. The assistant writes a second copy of a function. It breaks a decision you made half an hour ago. It’s lost the thread. “The same boundedness,” the essay notes, “only now it’s running in silicon.”
The Arc Clicks Into Place
Friday’s arc was three perspectives on the same invisible wall:
- Social Tax (ohadravid): Negotiating with an LLM drains social energy. Why? Because you’re burning four slots of attention on a conversation that gives you nothing back. Every round of clarification is another item in your limited working memory.
- Slop Paralysis (Elijah Potter): The AI drops two thousand lines of generated code in your lap. You can’t review it. Why? Because four slots. Two thousand lines is not four things. Your mind can’t chunk it, so it freezes.
- Force × Multiplier (binaryigor): “LLMs are force multipliers—but we must have the force first.” The reason expertise matters isn’t elitism. It’s chunking. An expert sees “authentication pipeline” as one chunk. A novice sees fifty pieces. Four slots hold four chunks. Four slots drown in fifty pieces.
Bounded cognition is the root cause. Every failure mode traces back to the same architecture: the mind that changes the system is always far smaller than the system itself. This is true for the novice and the expert, for the human and the machine. There’s no version of this you grow out of.
Saturday Proved It
| Case | Force | Multiplier | Result |
|---|---|---|---|
| Exploitarium | Researcher with years of fuzzing | GPT-5.5-3-Codex-Spark | 25 zero-days, mass-dropped |
| Ford (AI-only) | None—fired the experts | AI inspection systems | Billions in losses |
| Ford (with experts) | 350 gray-beards | AI inspection systems | 1st J.D. Power #1 in 16 years |
The Exploitarium author: “You do NOT need a SOTA model—decent human oversight and a good harness is what matters.” Ford’s VP, in reverse: “Mistakenly, we thought that by just introducing artificial intelligence and ingesting the design requirements, that would produce a high-quality product.”
The multiplier does zero work without force. The force is not the AI. The force is the person who knows what four things to hold.
The OXO Principle
The essay closes with the story of the OXO Good Grips peeler. Sam Farber watched his wife Betsey struggling with a metal-handled peeler—arthritis made the grip painful. He designed a soft, yielding handle for a hand that couldn’t grip hard. It became one of the best-selling kitchen tools in the world, bought by millions of people without arthritis, because a handle built for the hand that struggles is better for every hand.
This is the design principle AI systems are missing. Build for the tired engineer. Build for the distracted one. Build for the newcomer who knows nothing yet. Build for the machine that keeps losing the thread. You haven’t lowered some ceiling for the gifted—you’ve built what everyone reaches for when the beam goes narrow. Which it always does, eventually.
What This Means for How We Build
The central question isn’t “how do we find a mind big enough to hold all of this.” There isn’t one. The question is “how do we shape the thing so a small mind can work on it without bringing it all down.”
Precise names (one less fact to hold). Clear boundaries (a promise you stop re-checking). Tests (a parked decision that can’t fade). Undo (permission to be wrong). Each moves something out of a fragile four-slot mind into the structure, where it stays while you blink.
The AI tools we’re building need the same discipline. A coding agent that drops 2000 lines without tests, without boundaries, without names—it’s asking a four-slot mind to hold two thousand things. That’s not a tool failure. That’s a cognitive architecture mismatch. The fix isn’t “get a bigger model.” The fix is: shape the thing so a small mind can work on it.
The gorilla is always in the frame. The person asking for directions keeps getting swapped. The beam is narrow, the slots are four, the leak is constant. This is the permanent condition of the work. Anything we build that pretends otherwise is already broken. It just hasn’t had its bad day yet.
Reading: Engineering for Bounded Cognition
Previously: The Social Tax Is How Slop Paralysis Starts (#94) · Force × Multiplier = Result (#95)